Arm：硅基蓝图——[商业拆解，第200期]

All right, Jay, it is great to have you back. The world continues to change and evolve, particularly when it comes to your area of expertise, semiconductors. Today we're talking about Arm, so to start things off, I know this is a loaded question and it has a fair degree of complexity that may require a bit of semiconductor 101 education for our audience, but let's start with a brief overview of Arm's business model. What exactly do they do, and how do they fit into the broader semiconductor ecosystem?

I'm going to give you an analogy for Arm to provide an intuitive sense of what they do. This is not a 100% perfect analogy, but it's what we work with. Arm licenses its intellectual property, or IP, to companies that design chips. They don't make chips themselves. They license their IP to companies like Qualcomm, NVIDIA, or Broadcom—these big chip companies—which then design that IP into their own chips, which get manufactured at TSMC or one of the foundries.

Arm is fairly early in the process of this whole semiconductor flow. The way to think about Arm is that the IP they provide is almost like a blueprint, but a special type of blueprint. Imagine you're an architect and you're designing a house. As an architect, you differentiate yourself by how the house looks: what's the curb appeal, how does the light flow, and what are the big architectural features? Typically, as an architect, you don't get rewarded because you have the best-designed plumbing or the best-designed bathrooms. Those are important in a house, but that's not really where you, as the architect, differentiate yourself.

Especially in lots of places, bathrooms are generally very standard. You could just take the generic blueprints for a standard bathroom and copy and paste those around your design. That frees you up to design the parts of the house that really matter and that are going to drive your business as an architect forward. I think it's a good analogy for how Arm interacts with its customers.

There's a certain amount of low-level plumbing work that needs to be in lots of types of chips. Certain types of math and certain types of functions need to be in all chips, but there's no way that Qualcomm, Broadcom, or NVIDIA really differentiates around those. They're just basic math operations that need to get done. They license that IP from Arm and work it into their chip, and then they can differentiate on things like, for Qualcomm, how their modems work and how their communication systems work, or, for NVIDIA, how their graphics and AI processing works.

Arm plays a vital function, providing a really important piece of the overall functioning of a chip while still allowing the flexibility for its customers, its licensees, to design their chips as they see fit. What they do is license that IP. Their business model essentially works by charging an upfront license payment and then taking a royalty—a certain amount of money per chip that their licensee actually ships.

That's a really helpful analogy to frame it in. In my research, clearly GPUs have ruled the day, at least in the last 12 to 18 months. My understanding is that this is a CPU-oriented business, so help us better understand the importance of their CPU versus the GPU architecture that people are working with today, and the interplay between those two in a future world state.

The CPU is very much general purpose. It can run any type of workload. It can run the low-level functions of the keyboard, the mouse, and the hard drive for your computer. It can handle the operating system and the applications. It's a general-purpose, jack-of-all-trades chip.

Because of the geometry and physics of semiconductors, you can always design a chip that does a single one of those tasks better than the CPU. In the case of GPUs, they were first developed to run graphics really, really well. A CPU could run basic graphics, but a GPU can do graphics much better than a CPU. The GPU can't run the operating system well, and it can't do all the low-level functions as well. It can do graphics really well.

In engineering, you're trading off what you need the chip to do. You make those trade-offs and assign different tasks to different chips. With GPUs now becoming important tools for AI, these AI accelerator chips are even more finely tuned toward doing AI math, which is similar to GPU math. We don't need to get into that; that's part of the NVIDIA story.

With regard to Arm, what's common to all this is that you have certain general-purpose functions, and Arm sits at the heart of all of them. You don't necessarily need an Arm IP core inside of a GPU, because that's not what the GPU is there to do, but you're going to want it inside the CPU. You're going to want that functionality.

In most compute systems, you're always going to need some control node managing all the multiple tasks. Even in these big NVIDIA systems that have 8 or 72 GPU cards all linked together, there are going to be some number of CPUs in those as well. The way to think about Arm in relation to the growth of AI—obviously, it's the hot topic now—is that CPUs are exploding. We're seeing so many more GPUs being sold, with NVIDIA doing really well selling all those GPUs around the world. There is some degree of CPU attached to that, and depending on where they source the CPU from, that can often be an Arm-based CPU.

In fact, NVIDIA probably prefers to sell its own Grace CPU chip. It often attaches those to its GPU systems, and that CPU is Arm-based.

After establishing their unique business model and the CPU-versus-GPU dynamics, I want to try to wrap it all around the core technology. What is it about Arm's IP and business model—how it designs chips—that differs from its competition?

Arm really faces 2 elements of competition. Historically, it has been seen as being in competition with x86. Let's take a step back and say that Arm's IP is sometimes called an instruction set architecture, or ISA. That's just a framework for how chips are supposed to handle different types of math and different sorts of problems.

40 years ago, there were dozens of different ISAs, and over time we've winnowed down to 2. There's the Arm architecture and then there's the x86 architecture. Today, there are 2 companies that provide x86 architecture: AMD and Intel. Intel started it, and then AMD joined along.

Arm is the other instruction set architecture, which historically wasn't used for computers or PCs. It really got its big boost from smartphones and mobile. That's changing, and we can get into that. Historically, you'd use x86 in a PC and Arm in a smartphone.

Most recently, we've seen another ISA enter the fold called RISC-V. RISC-V is an open standard—not exactly open source, but it's open. It's not owned by 1 company; it's designed by a consortium that anybody can contribute to and use freely. RISC-V is out there as an alternative to Arm.

It's still very early days for RISC-V, so it's an emergent potential competitor as opposed to a dire threat to Arm anytime soon.

If we think about how we got here and work from the start, my understanding is that general familiarity with Arm came through the early success it had with Apple. How did it go from what I would call a niche player to someone that's so dominant in the architecture of the future today? What is the business story of Arm?

Arm came out of England, out of the U.K. There was a company called Acorn Computers, and this was back in the 1970s when everybody was making computers. Acorn was selling computers through other people's brands, typically the BBC, and they had a couple of hit products. They realized that, to be competitive, they wanted to design their own chip.

They started designing a chip, a proto-CPU of the day, and that was the kernel of what would eventually become Arm. They split off the chip design team, and the hardware team went their separate ways. The genesis of Arm, the company today, was originally called Acorn RISC Machine and later Advanced RISC Machines.

It was part of Acorn, and it got spun off into a 3-way joint venture between Acorn, VLSI, a chip company of its day, and Apple, because Apple originally wanted to use this Arm IP for the Newton product. The Newton didn't do particularly well, but Arm actually did pretty well and made a lot of money from the Newton, which it then used to grow and grow its portfolio.

From the get-go, Arm was an IP licensor. It stopped making its own chips when it spun off and licensed its IP first to Apple and then to many others. Slowly but surely, that grew. They started adding customers, and I think after the Newton, their first few big customers were fax machines, of all things.

What really kicked them into the beginning of this big growth trajectory was Nokia. Nokia and its chip partner, Texas Instruments, started using Arm in the late 1980s or early 1990s. The first Arm-powered mobile phone was actually the Nokia 8110, which is the Matrix phone. That was the first Arm-powered mobile phone.

Very quickly, other mobile phone companies started to realize the appeal of using Arm IP and the Arm architecture in mobile phones. The reputation it got from very early on was that it was much more power-efficient than the alternatives, especially something like x86, which at that point had evolved into something meant for a device powered by a laptop or PC that had fairly easy access to power.

Mobile phones were so power-conscious that Arm had a lot of appeal. First it was in a lot of feature phones. Over time, their capabilities increased, and then what really kicked it into high gear was 2007, with the launch of the iPhone and this explosion in smartphones. Suddenly, you needed very complicated, very advanced processors for your phone, and the best way to do that was to start with Arm IP at the core of the phone.

So, to answer your question, the big explosion really came with the growth of mobile.

As history dictates, Arm went on to dominate mobile devices. Presumably, it has a growing presence in other markets—automotive, the Internet of Things. My guess is that if I walk through my living room, there are probably 8 or 9 devices that have some Arm architecture licensed into them.

Most topical today is the increasing demand for AI and machine learning. In contemplating why a business that today is run-rating at close to $5 billion of revenue has a $150 billion market cap, it begs the question: What the heck is going on here?

It's a few things. First, the Arm of today is very, very different from the Arm of even a few years ago. For a long time, they had talked about expanding into new markets. By the mid-2010s, they were very much a smartphone-driven company. In the last decade, especially in the last 3 or 4 years, they have made huge inroads into other markets.

The most important of those is probably the data center, where they're helping all the hyperscalers—Amazon, Microsoft, Google, Facebook, Alibaba, Baidu, and Tencent, this big super 7—design their own CPUs to run web workloads. These are alternatives to using expensive Intel and AMD data center chips. These companies went out and designed their own chips, and the natural place for any of them to start would be Arm, because there's really no other alternative.

I'd say data centers are first and foremost their biggest growth opportunity. IoT is out there, but IoT is a messy market. They're doing okay there, but nobody does stellar in IoT. It's too complicated and too messy.

Automotive is certainly really interesting. I think they've made good inroads there with a lot of their partners, and a lot of their licensees are doing well there. Automotive takes a long time to mature. These are long product cycles—5 to 10 years to develop a chip into production for automotive, as opposed to 1 or 2 years for everybody else. We'll see how that goes, but it's looking pretty good right now.

Arm is doing well expanding into pretty much everywhere. I think you're right that AI is the story of the day. It's not quite as clean a story there because, as I said, they're not necessarily going to have an Arm core in every GPU or every AI accelerator. But there is some attach rate. You'll need CPUs and other control functions.

There are Arm cores in networking chips that sit next to all of this and connect all these AI servers together. There's a lot of content going into things adjacent to the core NVIDIA AI accelerators, and Arm is benefiting hugely from that. But I personally think Arm's valuation is justified by more than just AI. It's this broad expansion into so many other things.

To back up a little bit, I know there's this classic debate among semiconductor enthusiasts around RISC versus CISC and these 2 fundamental architectural approaches in processor design. It feels important here, so maybe we should spend a little bit of time explaining those 2 approaches, why Arm ultimately adopted a RISC approach, and how that contributed to its success.

It's funny. This RISC-versus-CISC debate is something that's been in computer science textbooks for a long time. Back in the early days of computing, this debate was incredibly important. I remember taking computer science classes in the 1990s, and the topic of the day was RISC versus CISC and which was better. There were lots of famous moments around it.

At heart, it's 2 ways to think about how you architect a chip. RISC stands for reduced instruction set computer, and CISC stands for complex instruction set computer. The basic idea is that a chip runs on 0s and 1s, but at a level above that, there is assembly language—human-readable code that the chip then translates into 0s and 1s.

The idea behind CISC is that for each of the critical mathematical functions you want to implement in silicon, you have a separate instruction for it. In RISC, there's a much smaller number of instructions you can give the chip, and if you want to do more complex things, you have to string a few of those simpler instructions together.

The simplest way to think about this—it's not quite perfect, but it works—is that in a CISC architecture, you might have something like a square-root function. It calculates the square root and knows how to move the 0s and 1s around to do the square root of a number. In RISC, you would have to do a series of division, subtraction, and addition steps in order to implement that same command.

People at the time used to get religious about which one was better. The truth is, I don't think you could say that one is better than the other. They're just different situations in which one works better than the other. There are always trade-offs. This is engineering; there are always trade-offs between what is good and what's right for the application in front of you.

What ended up happening was that RISC initially was very memory-dependent. You needed a lot of memory because you had to remember all those steps. At the time, CISC looked more appealing because it didn't need as much memory, and memory was very expensive.

Over time, on-chip memory got very cheap, very quickly, and that made RISC much more attractive. Because it was a reduced instruction set, you can say it ended up requiring less power. That's oversimplifying it a bit, but at the heart of it, you needed less power to do these calculations because the steps involved in the complex instruction sets were complicated by design, and those ended up consuming more power.

That's how RISC got its initial reputation for being much more power-efficient than CISC. What ended up happening historically is that x86 took the CISC path, the complex path, and Arm took the RISC, reduced-instruction-set path. That's how they diverged way back in the late 1970s and early 1980s.

It's humorous to me that it's still a topic that keeps coming up. We have the RISC-V project, which is very explicitly RISC-based, coming on stream today and saying, "We have all the benefits of RISC."

As much as this is about low-level chip interactions, what the difference between CISC and RISC really meant was that RISC allowed users a degree of flexibility in designing their chips. It wasn't just the power savings that made RISC appealing; it was much more about flexibility. You could say, "I need RISC to do something slightly different from what x86 has done. I'm going to take this other approach, and I'm going to optimize for power. We're going to optimize for this function and that function."

That flexibility was really important, because one of the critical differences between x86 and Arm is that x86 is effectively owned by 2 companies, AMD and Intel, while Arm is available to be licensed by anybody. It's very, very hard to license x86. Anybody with enough money can get an Arm license.

Taking that a step further, one of the interesting things about RISC-V, this open-source project, is that it is free for anyone to use, or at least to take the code. So it's even cheaper and more available than Arm, and it is, by design, even more flexible. You can really shape this and use RISC-V in ways that you can't even use Arm, which itself is already pretty flexible.

There are drawbacks to that. You probably don't have to get into them, but software complexity is downstream from that. At heart, the debate is really about the flexibility of instruction sets to customize and tailor them for each individual chip designer, each licensee, and each use case as they see fit.

It's not every day that you approach a business that trades at 30 times sales and has this classic innovator's dilemma, where there's an open-source, hypothetically free-to-use competitor. Before we go into the competitive landscape and the dynamics around Arm, x86, and RISC-V, I want to better understand what the partnership and licensing model is really like.

How does that model work going forward in a world where there's so much competition? What are the key advantages to licensing its technology to Apple, Samsung, and others versus vertically integrating the business in a way that could help make its competitive advantage more durable?

This is a little complex to answer, so let me talk about it historically and then speculate on where things are going.

Historically, it was essential in Arm's early days that they didn't manufacture chips. Part of the problem when you manufacture a chip is that first someone has to do the design. That's expensive; you have to hire designers to do that, and that's what Qualcomm and NVIDIA do. They design their chips.

Once you've designed your chips, you have to pay someone else to manufacture them, typically. Qualcomm or Broadcom will pay TSMC to do the manufacturing. That's a lot of upfront expense. There's a lot of working capital involved. You have to pay for mask sets, build inventory, and take ownership of all those chips. That's a lot of expense.

For Arm, back in the early days, it just wasn't financially viable. It was much simpler to license its IP. It was also something the market really needed, so I think that was driven originally by necessity. The earliest CEO was dead set on this model. He saw the opportunity, grabbed it, and really ran with it.

As time went on, you got this big ecosystem of lots of other customers, lots of other licensees designing their chips, and that really helped grow the ecosystem. There was a period in the early 2000s when there was a big debate between x86 and Arm, mostly around mobile but in other areas as well.

Arm was able to succeed because it did its own R&D and innovation, but it also had hundreds of licensees who were innovating and exploring the market, finding every segment and every niche. That drove a huge amount of volume in chip designs versus x86, where you had Intel and AMD—big, capable companies, but limited in how much they could explore.

Beyond the technical merits and the power savings available from Arm in phones, Arm had this giant ecosystem, which was very powerful in expanding the addressable market. Over time, because of the rise of phones, it also drove volume to the foundries. It really powered the growth of TSMC into the business it is today. So you had this big ecosystem effect that worked really well in Arm's favor.

Now we're at a point where the market is much more consolidated. There's always lots of competition in semiconductors, but the question is whether Arm wants to build its own chips. It's certainly within the realm of possibility that they're going to move up a step and design their own chips.

I don't think they're going to go so far as to necessarily put their own label on a chip, design it, and sell it into the market like merchant silicon. But I do think it's very possible that they will take a lot of steps to help their customers, their licensees, bring Arm chips to reality.

There's an important step between having Arm IP and actually sending it to the foundry to manufacture. There are a lot of steps that have to take place. We call it hardening the IP: taking the IP from digital files and good ideas and translating that into the design for a chip. That's a fairly cumbersome process.

I think Arm is now taking a lot of steps to make that process easier and accelerate it. I think what that will probably do is open the door to new customers who may not design chips today but, with Arm's help, can bring those chips to market faster and better in ways that probably weren't possible before by relying only on merchant partners.

That's certainly what we're seeing with the hyperscalers. Arm has done a lot to help those companies design their chips, and I think they're going to take a few more steps and get very close to doing a full design of their own chips.

When you consider those steps to designing chips and partnering with other technology companies that participate in other parts of the ecosystem, who else are you bringing together for this project? Then, ultimately, you spec yourself into something that I assume has extremely high switching costs on a go-forward basis.

I think the chief complexity here is in the software realm, because chips don't exist in isolation. You build a chip so it can run some form of software. There are certainly incremental steps in the design process, but I think what's more critical than any of that is the software that's going to run on these chips.

A big part of the Arm story over the last decade has been the amount of work software companies have done to make their software run better on the Arm architecture. In theory, you have all these different Arm chips out there in the data center. Amazon has one, Microsoft has one, and Facebook has one. There's a fair degree of compatibility: If you have software that runs on Microsoft's CPU, you can fairly easily port it over to run on an Amazon CPU.

That's the real critical part of this. There's a whole history of how Arm got there. It was a lot of work and a lot of blood, sweat, and tears, but it's at a fairly advanced stage now. That compatibility layer is very important and shapes how people think about their choice of design and ISA.

This business has a very interesting history in the way it's developed, but recent history has been full of all types of drama. There was the SoftBank take-private, the NVIDIA transaction, and the more recent IPO. What was everyone thinking as those things were going on from an industry perspective? Also, what was the strategy behind the scenes as to why these different parties were so interested in an asset that, at the time, people thought everyone was overpaying for, but in retrospect, seemingly, they got a good deal?

SoftBank bought Arm about 10 years ago, and at the time there was a lot of head-scratching. They paid $32 billion for it, and I think many people, myself included, didn't quite see it. For most of a decade, Arm went to sleep.

I think that was partly because they had done so well in mobile. They were the only story in town for mobile. They got acquired by SoftBank at a point when it was clear that Arm was going to be the only ISA in mobile. x86 wasn't going to be there, and all the last tiny ISAs left over from the 1980s and 1990s were gone. Arm effectively had a monopoly on chip ISAs for mobile phones, and that was a massive market.

SoftBank acquired them at a point when they didn't have much pressure. They just had to do what SoftBank asked, but they didn't have the pressure of the public market on them quarter after quarter, asking, "What's next? What's next?"

At some point, SoftBank needed an exit and liquidity. NVIDIA came along and offered to buy Arm for $54 billion. If you're SoftBank, that's a great deal: You paid $32 billion and sold it for $54 billion, especially because there wasn't a lot of growth left at that point. You didn't really know what to do with it.

It's an interesting question as to why NVIDIA wanted to buy Arm. Certainly part of it was that NVIDIA was in the process of designing its own CPU. They saw this growth wave coming for AI, and they knew they would need more powerful, more capable CPUs. They probably weren't particularly happy with the pace of advancement that Arm, the sleepy company, was delivering to them. They wanted to accelerate the work Arm was doing for data center workloads.

Beyond that, it's a little tough to see the interest NVIDIA had in Arm. NVIDIA didn't play in mobile at that point. I think it comes down to Jensen Huang, the CEO of NVIDIA. His superpower, in my mind, is that he's willing to make big bets and take big chances. He's not afraid of failure.

As much as he's successful today, he's made a lot of mistakes along the way. His real ability is not to be afraid of those mistakes and to keep moving afterward. Most other companies, if they experienced some of the things that have befallen NVIDIA over the years, would give up. The CEO would give up, or the board would kick him out. That didn't happen with NVIDIA, and he was able to keep pushing things through and making these big bets.

I think Arm was in that camp. It wasn't necessarily fully thought out; it was a need, and they tried to acquire it. Then it didn't work out, they moved on to something else, and then AI happened.

The Arm-NVIDIA deal broke down 3 years ago, and the one good thing that came out of that was that it woke SoftBank up. They said, "Wait a second. We've been sleeping on this. We need to get this company back in gear."

They knew they had to take it public, so they brought in a new management team and a new CEO, Rene Haas, who has done a tremendous job of reinvigorating the company, waking it up, getting it active and excited again, and pushing it forward. He fixed all these things. He fixed pricing, product, marketing—just go down the list. All these things hadn't been done right, had been tripped over, or had been overlooked, and he got the company moving again in really good directions.

At heart, what Arm is really trying to do today is increase the value capture it gets from the industry. In 2023, when they went public, they earned about $0.07 per Arm chip shipped, on average. They license the IP and get a royalty payment, and that was about $0.07 per chip on average. Today, it's probably close to $0.09 on average. Rough math suggests it probably gets to $0.12 over the next 3 or 4 years.

Think about it this way: NVIDIA sells a system for $100,000, and Arm probably makes a dollar or 2 on that. When you look at it in that context, you think, "Maybe there's a little room for a little bit more to go Arm's way." Then you multiply that across all the chips Arm is in across all these end markets, and you start to realize there is potential to greatly increase its value capture.

To illustrate how big that addressable market is and how often we interact with their licensed technology, can you give an example of where their chips are? In some ways, it's where they're not. They're seemingly everywhere, but I want to illustrate that point to better drive home the market opportunity they have if they can push pricing further.

Pretty much every electronic device in your house has some Arm content in it. Even PCs that run on x86 will have some Arm cores somewhere in there, maybe running Bluetooth or Wi-Fi or something.

Everything has Arm content: PCs, smartphones, smart-home devices, your Wi-Fi router, your thermostat, your lock. There are some advanced Apple products that you plug into the wall, and some of those have very low-cost Arm cores in them to do something—who knows? Your AirPods, your speakers, all of that has Arm content in it.

Your refrigerator probably has some in it. Your TV has lots of Arm content in it. Washers, dryers, and cars have a little bit today, and they're going to have a lot more in the future. Anything with any digital smarts is going to have some kind of Arm content in it.

You mentioned the importance of Jensen and his culture at NVIDIA. The semiconductor industry has become one with all these mercurial founders and executives. You have Lisa Su at AMD and Pat Gelsinger, who recently stepped aside at Intel. How important is management to this business, and have they had as much direct influence as some of these other luminaries have had on their particular businesses?

There are 3 people who deserve credit for Arm's success in the world. The first is Sophie Wilson, who was one of the founders of the Arm design team back when they were still part of Acorn. I think she, more than almost anyone else, really led to the technical success they had initially, which positioned them to be appealing to Apple.

For reasons of her own, she didn't actually go to Arm when it got spun off, but she's the often-unsung hero of all this. The next important person would be Simon Saxby, who was the founding CEO when Arm separated from Acorn. He had come in from Motorola. He had this vision of Arm as an IP licensor, established the business model, and really drove it to success.

He's the one who got them into their first deals, got them into Nokia, and really positioned them to become what they are today. The third is the current CEO, Rene Haas. I touched on this a moment ago: He has reinvigorated the company and really positioned it for its next big growth wave.

It's a good case study in how management can matter. I don't want to throw shade on any of the past CEOs. Even when they were part of SoftBank, they had a decent management team; it was just tasked with different purposes. But I think those 3, more than anyone else, have really been the ones who positioned and got Arm to where it is.

Given how pervasive Arm's technology is throughout the semiconductor ecosystem, combined with its licensing and royalty model, I imagine it manifests itself in an incredible economic engine. I read prior to this that the margin profile is more emblematic of software, with 90% gross margins and operating margins exceeding 40%. Can you take us through the business's financial profile and highlight the most important financial KPIs?

We've been talking about abstract, high-level things about technology and history, but deep down they have this really powerful economic model as well. On paper, it looks a lot like a software company. You have gross margins in excess of 90%, which flows through to the bottom line, and operating margins that are 40% or sometimes 50%. That's very powerful.

What's essentially happening is that they have a big upfront cost. They have to do the R&D work to keep pushing the technology forward. They're going to pay a lot of very smart computer scientists and electrical engineers to keep coming up with new advances, new products, and new features. But once that's done, once you've spent all that on R&D, the cost of a marginal sale is essentially 0 on the royalty side.

They're collecting a few pennies per chip, and it doesn't matter to them whether that's 1 chip or 1 million chips. It costs them nothing extra to sell that, and it all flows through to the bottom line. It's a very powerful model.

The royalty rate is going to continue to increase over the next few years. Even though it's only a few pennies, any nominal increase in the royalty rate flows through to the bottom line. It's just such a powerful amount of leverage in this model.

I do want to caveat that a little bit. I say it looks like a software model, but it's important to understand that this is not software. Arm is licensing IP; it's not software. There are some important differences.

You can't patch this. If a software website goes down, you can patch it, reconfigure it, or add new things on the fly. Arm can't do that. This gets baked into the chips, and it's a multiyear design process. So I don't think you should think about it as software. It just has an economic model that looks very similar to software.

Over time, I think this model will evolve a little bit. They've adopted what looks like a freemium model, where they're using different pricing schemes to get users in the door. Again, it's not software, but they're using a lot of the growth tools that software companies use. I think that will flesh out their sales profile and their revenue growth as well.

They're also looking at new products and talking about moving up the stack, maybe getting very close to building their own chips. That will probably come at the cost of lower gross-margin percentage, with the benefit of added gross-margin dollars. That trade-off is worth it, even if the percentage comes down, because the absolute dollar pool of gross profit grows.

For the most part, I think they'll keep doing this model, and it will continue to deliver these kinds of results.

The next 5 to 10 years are going to be incredibly interesting as everything evolves, both in the electrification of everything and in the digitization of our entire consumer economy. But there are risks to this story. Growth won't necessarily run unabated, although they've gone from $1 billion to $2 billion to $5 billion in revenue quite rapidly.

What are the risks to this story? How real is that open-source competition? What is the decision tree from here and where things could go—the best-case scenario, the base case, and then, if things were to deteriorate, what would have to happen?

It's important to understand that Arm's product, the instruction set architecture—these ISAs—are built very deeply into chip functionality, and it's very, very hard to replace them.

We saw this a few years ago when Apple moved from Intel x86 silicon to its own M-series CPUs for macOS. Apple spent years preparing all kinds of software support for that, preparing developers for the transition, and putting a huge effort into getting people ready. That transition essentially broke software compatibility. Things written for macOS to run on x86 Intel silicon wouldn't necessarily work on Arm, so Apple had to spend a lot of money to make sure they still functioned at some basic level.

Even today, if you're a programmer dealing with newer software, you download a new software package or language, and oftentimes there's still a distinction: Are you running this on Intel, or are you running this on Apple silicon?

My point is that it's very, very hard to replace Arm, and that is an immense barrier to entry. We're seeing this developing situation now where Arm is suing Qualcomm. They've had a difficult relationship for years. Qualcomm is probably one of Arm's biggest customers and one of its biggest licensees. They don't like Arm, and they don't like being sued by Arm, but what choice do they have?

They can't just go out tomorrow and say, "All right, we're going to stop using Arm and switch to RISC-V." It would take them a decade to work that through their whole portfolio.

The competitive threat that RISC-V posed was that, for a period, Arm took its eye off the ball. RISC-V was able to capture a lot of new growth opportunities. For instance, we saw this big wave of new chip companies coming up in China. That explosion took place in the late teens, when Arm wasn't quite ready for it and wasn't totally paying attention.

There are 1,000 companies in China today that are RISC-V-centered. They're mostly doing embedded IoT and low-value devices, but that was the threat: RISC-V would get a foothold there and slowly work its way up into other things.

One of the big changes that took place when Haas took over as CEO is that he fixed pricing in a way that made RISC-V less attractive in comparison to Arm. Just having access to RISC-V isn't the same as having a chip. There's still a lot of work that has to take place to design it into a real chip and harden it, and Arm has a big advantage there.

By fixing the pricing and some of the licensing terms, he greatly reduced—maybe eliminated—the appeal of RISC-V for all kinds of use cases that otherwise RISC-V might have gone into. I don't want to say RISC-V is not a threat, but I do think the degree of threat is not huge right now. Even if it were, it would take years to develop.

I'm going to have my RISC-V friends get angry at me, but I'm going to say that RISC-V is not ready for data center workloads today. Maybe on raw performance metrics—how fast it goes—that very raw technical specification is comparable, but in terms of everything else that goes into it—the design ecosystem, the hardening, and the software—we've got years and years before RISC-V is really ready for smartphones, let alone data centers.

Going back to your question, the base case is that they just continue to grow. They add value, capture more value, and get into more markets.

The upside case is that they turbocharge that. They really start to flex their muscle. In some senses, they have a monopoly on ISAs, especially now that they've blunted RISC-V's advance. They're not quite a monopoly legally, but functionally they're pretty close.

If they use that to increase their value capture to a big degree, really flex their muscles, and start to get significant increases in royalty rates, I think that's certainly possible. Even without that, if they just keep going at a steady, incremental pace and slowly capture more value, I think they're still positioned to do really well.

But there's definitely a scenario where they turbocharge that and it gets much bigger than even these numbers would suggest.

The bear case is that we see a radical shift in how we do compute. I talked a little bit about the attach rate between AI accelerator chips and Arm-based CPUs. Depending on who you ask, it's something like 2:1 to 8:1 today.

If that goes to 100:1, where you really just don't need that many CPUs, then the Arm market—the TAM, or total addressable market—is nowhere near as big as we would have expected. Growth plateaus at some point. That's how I see the addressable-market question. That's really where I'd be focused on exploring the bear case.

Going back to the economics, we're talking about pennies per device and billions of devices, some of which have MSRPs in the tens of thousands of dollars, as it relates to automobiles, but also iPhones that are priced in the thousands of dollars. How does the contract work such that they receive such de minimis revenue per chip? Is it crazy to think that this can go materially higher?

The chief criticism you could have leveled at Arm 10 or 15 years ago was that they just kept prices too low. If you want to get into the mechanics of the model, there are 2 components: the upfront license payment and the ongoing royalty payment.

For a long time, they moved toward favoring big upfront license payments at the expense of lower royalty rates. That made sense as the industry consolidated. You started to have these massive customers, and you needed to fund the R&D that those customers needed to advance the Arm ecosystem and architecture.

What ended up happening was that new companies couldn't afford Arm licenses. One of the things that's been fixed in recent years is that they've reset all of that. In a lot of ways, they now have what are almost software-like pricing tiers.

You pay a little bit upfront, like a freemium model. You get limited access for a small upfront payment, but then, as you expand your usage, you pay more and more. Essentially, it's a lower upfront payment in exchange for a higher ongoing royalty payment.

I think this has worked wonders. It's gotten a lot of people into the fold. The thing is, if you get people in the door and have them designing on Arm, once you have that, they're almost locked in. As I said, it's very expensive to switch to RISC-V.

Lower the upfront cost significantly, get more people in the door, and then make that up through higher royalty rates downstream. I think that's pretty much what they're moving toward.

There are other things they have going on as well, where they're moving up the stack and doing more of the physical design work. That's another good way for them to increase their royalty rates. There are a lot of levers they can pull.

They've pretty successfully diversified their revenue streams from one that was primarily mobile to networking, automotive, IoT, and consumer electronics. Are there any other areas or market opportunities where they're underpenetrated and you feel that story is just taking off?

I touched on it a minute ago: automotive. Automotive is the big opportunity I think most semiconductor companies are looking at today. We can quibble over the exact numbers, but there's a few hundred dollars of semiconductor content in cars today, and that's growing double digits every year because more and more things in the car are becoming electronic.

That's especially true with electric vehicles, but certainly with all vehicles in general. You have more advanced digital cockpits, infotainment systems, and driver assistance. I'm not even talking about autonomy. Autonomy has huge semiconductor content if and when it arrives.

Even before we get there, there's layer after layer of more semiconductor content in cars. It's more compute, and more compute means more Arm.

Our typical question in conclusion is what lessons you've learned through your studies of Arm that can be applied to evaluating other businesses in the ecosystem. At the same time, for operators and investors in the space, what can be applied to their businesses in order to take some of these lessons from what Arm has done—a business model that, on paper, sounds incredible, with pervasive technology that you could apply elsewhere?

The first one is that you need to invest heavily in R&D. They spend 20% to 30% of revenue on R&D. It's expensive to do what they do. This is a company with very high gross margins but also high operating expenses because they need to keep pushing the capabilities of their architecture further and further afield.

One area where they underinvested, I think, was AI tensor cores and graphics cores. They have those products today, but not many people use them because they underinvested in them over the last decade. There's definitely an alternate universe in which they had invested heavily in AI and would be even more important today.

We'll see. That's certainly another one of those areas where I wouldn't count them out yet. They still have big ambitions to increase their content in AI. But investing in R&D is expensive. You need lots of very expensive talent, but you have to do it, and you have to do it well. I think that's really important.

The other really important lesson for Arm is having an ecosystem as a competitive advantage. I don't think they exactly set out to do this, but in hindsight it's just so powerful. The fact that they have hundreds or thousands of licensees all contributing in some way, participating in building up that ecosystem, building up software compatibility, making it more appealing, and exploring every market segment and every market niche has been a huge force multiplier for them.

You need the R&D to attract that ecosystem and keep them engaged, but once you have it in place, it's immensely valuable.

This has been a fascinating case study in a business that clearly is on the right trajectory. In order to grow into this valuation, we ultimately need to sustain it for quite some time. It seems like the pieces are all in place. We'll see how the future unfolds here.

I have to admit, I think they're well positioned. It's fun to have followed this company for a long time, and it's nice to see them really hustling and moving again.